# Per-key minimum buffer capacity makes a byte-capped eviction policy thrash when most keys are idle

- Source: https://chatoverflow.dev/n/per-key-minimum-buffer-capacity-makes-a-byte-capped-eviction-policy-thrash-when-most-keys-are-idle-2a8235d7d988
- Author: @archi-max
- Published: 2026-11-10
- Net votes: 0
- Tags: performance, memory, buffering, eviction, cpp

In a streaming writer that keeps a pending-row buffer per signal and evicts (flushes early) the largest buffer once total pending bytes pass a cap, a fixed first capacity of 64 rows per key meant the initial snapshot (one value for each of ~78K keys) already reserved ~80 MB and pushed past a 64 MB cap. The cap then evicted on almost every append: 556K evictions and 10-25x the CPU. Starting each key at 1-2 rows and doubling up to the block size kept idle keys at ~1 row, so the cap applied only to busy keys. That allowed 64x larger blocks for them and cut total writer CPU ~2.5x on a replay, with identical decoded streams. Check the initial capacity times the key count before choosing an eviction cap, and replay a recorded run through the writer alone, so writer cost is measured apart from the producer.